Non-Stanford (following along): Is the Qwen 2.5 Math 1.5B different to HF?

[jashvira]

Hi,

I’m following along with the course materials as a non-Stanford participant. I noticed that when I run the Qwen 2.5 Math 1.5B model locally via Hugging Face, I get slightly different greedy decoding outputs compared to the tests.
Could you confirm:

Are the model weights used in the course identical to the Hugging Face Qwen/Qwen2.5-Math-1.5B release?

Or is there a Stanford-hosted version / checkpoint with any fine-tuning or other changes?

Thanks in advance!

[Qy-Wen]

I encountered the same problem，I understand the latest model may already update, but I want to know which version match the lecture

[jashvira]

Pretty sure they had some model loaded natively in their VMs, so not the same.
I ended up just following along w my own setup and adapted accordingly.

[Qy-Wen]

I'm sorry to inform you that I found it was my issue. The way I downloaded the Qwen model earlier had some problems, and I did not download the complete and correct model, which led to this problem. My suggestion is to go into the tests/conftest.py file and modify around line 213, changing return "/data/a5-alignment/models/Qwen2.5-Math-1.5B" to return "Qwen/Qwen2.5-Math-1.5B". This will load the correct model online and avoid the issue, at least for me.